Record meetings on your Mac.
Keep them on your Mac.
Your transcripts. Your AI. Your destinations.
Daisy records meetings locally on your Mac, transcribes them on the Neural Engine, then exposes them as a local MCP server Claude Desktop and Cursor can query — or pushes them to Notion, Linear, Attio, or a webhook. Nothing leaves your machine unless you say so.
Free during beta · Lifetime after launch — never a per-meeting subscription
Apple Silicon (M1+) · macOS 14+
Not only meetings
Hold a key, talk, and it lands as text — in any app.
Daisy isn’t only a meeting recorder. The same on-device pipeline powers push-to-talk dictation: hold your hotkey, speak, and the words appear at your cursor wherever you’re typing — with nothing leaving your Mac.
Works anywhere you type
Email, Slack, your editor, a form field — hold the hotkey, speak, release. The text is pasted at your cursor via the Accessibility API, and your previous clipboard is put back afterwards.
On-device, your choice of engine
Whisper on the Neural Engine by default, or switch to Parakeet (FluidAudio) for lower latency. No cloud round-trip and no API key needed — dictation never leaves the Mac.
Teaches your words
A vocabulary dictionary fixes names, brands, and jargon the model would otherwise mishear, applied right before the paste. A rolling 24-hour history lets you re-copy anything you dictated.
Live data source
Your transcripts, available to Claude Desktop and Cursor — without ever leaving your Mac.
Daisy ships a local MCP server bound to 127.0.0.1. Flip it on, click Add to Claude Desktop, and your meeting history becomes a queryable — and actionable — data source for any AI client that speaks MCP: read any transcript, then re-summarize it, name the speakers, or route it to Notion or Linear. No copy-paste, no API token, no upload — the data path is your Mac talking to your Mac.
The tools your AI can call
list_sessionsDiscover what's been recorded — title, date, duration, folder. Metadata only, no transcript bodies leak through.get_sessionPull the full content of one recording: transcript, summary, action items, attendees, timestamps.search_sessionsSubstring search across titles, transcripts, and summaries — your AI finds the meeting on its own.rename_speakerTell it 'speaker A is Maria' — the transcript updates and Daisy remembers the voice for future recordings.route_session_to_destinationPush a finished session to Notion, Linear, Slack, or a webhook — the same Send-to action as in Daisy's UI.
Nine tools total — five read, four act (re-summarize, retitle, rename speakers, route). Action tools are scoped to safe, reversible operations: no deleting, no editing transcripts.
{
"mcpServers": {
"daisy": {
"command": "npx",
"args": [
"-y", "mcp-remote",
"http://127.0.0.1:54321/sse",
"--transport", "sse-only",
"--allow-http"
]
}
}
}One click writes this into ~/Library/Application Support/Claude/claude_desktop_config.json — preserving any other MCP servers you already have. Claude Desktop speaks stdio, so a tiny mcp-remotebridge proxies to Daisy's local SSE server. Restart Claude Desktop, and your transcripts are live.
Straight talk
Honest about “who said what”.
Speaker labeling is the part every meeting tool quietly gets wrong. We’d rather tell you how it actually works — and how Daisy is built to land on the right side of the numbers.
The words are the easy part
Modern speech-to-text gets the words right ~95% of the time. Knowing who said each line — diarization — is the hard part, and it trips up every meeting tool. On real meetings even the major cloud engines land around 26–27% diarization error, within a point of each other. Anyone showing you a spotless speaker score is hiding the cross-talk.
Daisy re-runs it offline, the moment you stop
Real-time diarization is far worse than offline — often 26–50% error, splitting one person into several. So Daisy doesn't trust the live pass: when you hit Stop, it re-diarizes offline. The on-device engine it uses (FluidAudio, from the pyannote family) scores ~10–15% error offline on standard benchmarks — on par with pyannote itself. Your saved transcript gets that grade, not the streaming one.
Two channels, not one guess
Your microphone is always you. The other side of the call is captured on its own channel. So diarization only has to sort voices within the remote side — it never has to guess which stream is you. Fewer ways to be wrong.
When it's wrong, fixing it is one click
Overlapping speech and cross-talk still defeat everyone, us included — so we made the fix trivial. Rename a speaker once, and Daisy remembers their voice and labels them automatically in every recording after.
These are published third-party benchmarks, not numbers we cooked up — we haven’t run our own labeled benchmark yet, so we won’t pretend to. Sources: Cloud DER on real meetings (Scribie) · FluidAudio offline + streaming DER · pyannote accuracy · Why cross-talk is the limiter (Circleback)
The deal
Your meetings stay on your Mac.
Every other meeting tool wants your transcripts on their server — for indexing, for AI training, for whatever they haven’t told you yet. Daisy doesn’t. Here’s exactly what that means.
Transcripts live in your folder
Daisy writes Markdown into the folder you pick — typically an Obsidian vault or your iCloud Drive. Inspect, copy, delete, version-control: they're plain files on your disk. No "feature" that uploads them "for your convenience".
~/Obsidian/Daisy/Sessions/Transcription runs on the Neural Engine
WhisperKit runs fully on-device. Your audio is decoded into text by your own Mac — the same chip that does Face ID and Live Text. We never see it.
Summaries — your call, your key
Apple Intelligence works fully offline. If you bring an Anthropic or OpenAI key, that traffic goes from your machine straight to their API. Daisy is not a proxy. We never see your meetings or your key.
MCP server is bound to localhost
When you flip on the MCP server so Claude Desktop or Cursor can read your transcripts, Daisy listens on 127.0.0.1 only. Other Macs on your Wi-Fi, your phone, your work VPN — none of them can reach it. The server stops when you flip the toggle off.
http://127.0.0.1:54321/sseNo telemetry. No tracking. No account
Daisy doesn't phone home. There's no signup, no email, no "pro plan" that unlocks if we know who you are. You install. It works.
Bring your own AI
Pick the brain. Daisy’s just the wiring.
We don’t lock you into one provider. Use the model you already trust — pay them directly, or run it yourself on the same Mac that’s recording.
Cloud
Bring your own key
Plug an Anthropic or OpenAI API key into Settings. Your transcript goes from your Mac straight to the provider — Daisy isn't a proxy, doesn't see the key, doesn't see the prompt.
Anthropic Claude
OpenAILocal
Run it offline
Point Daisy at Ollama, LM Studio, or Apple Intelligence — or any local model behind an MCP server. Zero network calls. Apple Intelligence works without setup; Ollama and LM Studio are a one-line install and keep your meetings entirely on-device.
Ollama
LM Studio
Apple IntelligenceFor teams & enterprise
The compliance review is short, because there’s nothing to review.
Daisy was designed for the meeting-tool category that can’t exist in a sanctioned-data environment. Transcripts stay on the laptop they were recorded on. There is no Daisy cloud to audit, no third-party subprocessor list, no vendor pipeline carrying your customer conversations to a server you don’t own.
No DPA to sign
There's no Daisy server holding your transcripts, so there's no data processor to designate. Your IT counsel doesn't need to negotiate terms with us — we're not in the data path.
Your AI vendor, your contract
Daisy is BYOK across Anthropic, OpenAI, Apple Intelligence, and any local model (Ollama, LM Studio, or an MCP server). If your company already has an Anthropic enterprise contract or a self-hosted LLM, summaries run on that. No second AI bill.
Open source
Daisy ships under Apache 2.0 with full public source on GitHub. Your security team can read every line, build it from source, and verify there's no telemetry — instead of taking our word for it.
Mac-native, signed, notarised
Hardened Runtime, Apple Developer ID signed, notarised by Apple, distributed as a DMG. MDM-friendly. No Electron, no third-party update channel, no Chromium runtime to patch.
Get Daisy
Open source, on your Mac.
Daisy is published on GitHub under Apache 2.0 — the source is there, the signed builds land on the Releases page as they ship. No email collection, no “we’ll let you know”.
Apple Silicon · macOS 14+
Built by Addicted Studio. Native Mac app. No cloud. No login. No subscription gate.